SADL

SADL (Size-Adaptive Decision Layer) was proposed for the ISLES 2026 challenge, combining ensemble learning with a post-processing decision layer for chronic ischemic stroke lesion segmentation.

The underlying architecture of the SADL solution is an eleven-member ensemble model based on nnU-Net, differentiated across five design axes.

The five design dimensions of SADL include: a baseline loss function combining Dice and top-K losses, strong data augmentation, reweighting by lesion volume, Stochastic Weight Averaging (SWA), and the pairing of cohort-balanced sampling with a residual encoder backbone.

The SADL decision layer executes an adaptive process: estimating lesion volume and applying size-bucketed thresholds, pruning connected components by voxel count and peak probability, and incorporating a 'never-empty' rescue step.

The never-empty rescue mechanism is especially critical in small-lesion scenarios, intercepting small-lesion samples whose probability falls slightly below the threshold and would otherwise be silently discarded.

All fusion weights and bucketing parameters of SADL are fitted on out-of-fold (OOF) predictions, and the tuning process does not touch the held-out test set in order to avoid data leakage.

On a site-disjoint held-out set of 91 cases, SADL achieved a Dice coefficient of 0.7322, a lesion-level F1 of 0.7461, and a PR-AUC of 0.8430.

The researchers used a paired non-parametric bootstrap to compare SADL against a reference single-member ensemble, confirming that the improvement in lesion-F1 is statistically significant (p=0.024).

The largest gains from SADL occur in small-lesion scenarios, validating the targeted design of the never-empty rescue decision layer.

SADL links pruning thresholds to lesion size buckets, applying relatively lenient retention criteria for small lesions and relatively strict fragment penalties for large lesions.